Frailty Models for Heterogeneous Event Risk

Why Model Unobserved Differences in Event Risk?

Patients with the same measured covariates and drug exposure may still differ in their underlying susceptibility to an event. Some may be more vulnerable because of unmeasured disease severity, physiology, or other risk factors. A frailty model represents this residual heterogeneity with a positive latent multiplier on the hazard. It can change not only predicted event probabilities but also the relationship between a conditional exposure effect and the effect seen after averaging across patients.

The key distinction is between the event process conditional on frailty and the process marginal over frailty. A proportional hazards model may hold at the conditional level while failing to hold at the marginal level. Shared frailty also creates dependence among people in the same cluster, even when their event times are independent conditional on the cluster frailty.

Model Specification

For individual \(i\), let \(C_i(t)\) denote drug concentration or another exposure measure at time \(t\), and let \(W_i\) denote additional measured covariates. A conditional hazard specification is

Here \(h_0(t)\) is the baseline hazard, and \(\beta\) is the exposure coefficient conditional on frailty and the modeled covariates. This expression assumes the exposure effect is log-linear on the hazard scale; other exposure-response functions can replace \(\beta C_i(t)\). The hazard must be nonnegative, and the cumulative hazard must be finite over the time interval under consideration.

Define the cumulative hazard without frailty as

This definition allows concentration to vary over time. Conditional on \(Z_i=z\), survival through time \(t\) is \(S_i(t\mid z)=\exp\{-zH_i(t)\}\). The frailty is assumed constant over the modeled follow-up for that individual, positive, and independent of the measured covariates unless the model explicitly specifies otherwise.

Marginal Survival: Integrating Out Gamma Frailty

To obtain population survival, average conditional survival over the frailty distribution. For the stated gamma parameterization, the Laplace transform is \(E[\exp(-sZ)]=(1+\theta s)^{-1/\theta}\) for \(s\geq0\). Therefore,

The formula is not the same as substituting the mean frailty into the conditional survival function. In general, \(E[\exp(-ZH)]\neq\exp[-E(Z)H]\). The marginal event risk by time \(t\) is \(1-S_i(t)\), whereas the conditional event risk for an individual with frailty \(z\) is \(1-\exp[-zH_i(t)]\).

A useful consequence follows by differentiating the marginal cumulative hazard \(-\log S_i(t)\). The marginal hazard is

The denominator grows with accumulated hazard when \(\theta>0\). At later times, high-frailty individuals are more likely to have experienced the event, so those who remain event-free are increasingly selected toward lower frailty. This selection changes the marginal hazard even though each individual's conditional hazard still follows the specified model.

Individual and Shared Frailty

An individual frailty assigns a separate latent multiplier \(Z_i\) to each person. If frailties are independent across people, this model represents person-level heterogeneity but does not, by itself, induce dependence between different people's event times. If the same individual's frailty multiplies the intensity for multiple recurrent events, however, it links that person's event history: a person with high \(Z_i\) tends to have higher event intensity across that person's events.

A shared frailty assigns one multiplier \(Z_g\) to everyone in cluster \(g\), such as patients in a hypothetical site or family. Conditional on \(Z_g\) and measured covariates, the members' event times may be modeled as independent. Marginally, they are dependent because they share the same unobserved risk multiplier.

For example, if the cumulative hazards without frailty for members \(1,\ldots,m\) in a cluster are \(H_1(t_1),\ldots,H_m(t_m)\), their joint survival under a shared gamma frailty is

This is generally not the product of the marginal survival probabilities. Consequently, treating clustered event times as independent after integrating out a shared frailty can understate their dependence.

How Frailty Changes the Exposure Effect

At a fixed time, with the same baseline and covariates, let \(r=\exp(\beta\Delta C)\) be the conditional hazard ratio for an exposure contrast \(\Delta C\). For a simple constant-exposure comparison, write \(H_0(t)\) for the unexposed cumulative hazard without frailty. The exposed cumulative hazard is \(rH_0(t)\). The marginal hazards are then

Their marginal hazard ratio is

When \(r>1\), \(\theta>0\), and \(H_0(t)>0\), this marginal ratio is less than \(r\). It can also vary over time because cumulative hazard changes over time. Thus a constant conditional proportional-hazards exposure effect does not generally imply a constant marginal hazard ratio. The direction and size of the contrast depend on the exposure effect, frailty variance, and accumulated risk.

Assumptions, Estimation, and Model Checks

Frailty models require substantive assumptions. The frailty distribution is a chosen description of latent heterogeneity, not a fact revealed directly by the data. The mean-one convention separates the overall hazard scale from the frailty multiplier; \(\theta\) controls variation around that mean. The exposure and baseline hazard forms must also be adequate, and exposure measurement and covariate adjustment must support the intended interpretation. A conditional exposure association is not automatically causal.

For a single event observed at time \(t_i\), let \(\delta_i=1\) denote an event and \(\delta_i=0\) denote right censoring. Conditional on frailty, the event-time contribution has the form

where \(h_i^*(t)\) is the hazard without frailty and \(H_i(t)\) its cumulative hazard. A censored observation contributes conditional survival rather than an event hazard. For an individual frailty model, the conditional contribution is integrated over that person's frailty distribution. For a shared frailty, the product of member contributions is integrated over one common cluster frailty. These likelihoods rely on censoring being noninformative conditional on the modeled quantities; event-dependent dropout requires additional modeling.

The gamma distribution yields the closed-form marginal survival shown above, which is one practical advantage. A lognormal frailty is a common alternative, but typically requires numerical integration. A model without frailty is simpler and may be adequate if residual heterogeneity is negligible; it cannot represent the latent risk variation or shared dependence targeted here. A shared-frailty model is appropriate when cluster-level dependence is scientifically plausible, while an individual frailty is suited to persistent unobserved differences between persons. These structures are not interchangeable.

Useful checks include examining whether estimated \(\theta\) is near zero or poorly determined, whether conclusions change under a plausible alternative frailty distribution, and whether the chosen exposure and baseline hazard forms produce reasonable predictions across follow-up. In shared models, the assumed clustering structure should match the study design. Weak information about long-term event risk, few clusters, or limited events can make the frailty variance difficult to estimate. A numerical estimate of \(\theta\) should not be treated as proof that the latent risk distribution is gamma.

Common Pitfall and Practical Interpretation

Frailty matters in PK/PD because patients can have different underlying event susceptibility even at similar measured concentrations. Ignoring this heterogeneity can distort population survival predictions and, in clustered settings, fail to represent dependence. A frailty model does not remove the need to specify the exposure effect carefully: the conditional model describes effects given latent risk, while marginal predictions average over the modeled frailty distribution.